2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2010,2011,2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
40 #include "gromacs/utility/smalloc.h"
41 #include "types/hw_info.h"
43 #include "gpu_utils.h"
44 #include "../cuda_tools/cudautils.cuh"
45 #include "memtestG80_core.h"
47 /** Amount of memory to be used in quick memtest. */
49 /** Bit flag with type of tests to run in quick memtest. */
50 #define QUICK_TESTS MOD_20_32BIT | LOGIC_4_ITER_SHMEM | RANDOM_BLOCKS
51 /** Number of iterations in quick memtest. */
54 /** Bitflag with all test set on for full memetest. */
55 #define FULL_TESTS 0x3FFF
56 /** Number of iterations in full memtest. */
59 /** Bit flag with type of tests to run in time constrained memtest. */
60 #define TIMED_TESTS MOD_20_32BIT | LOGIC_4_ITER_SHMEM | RANDOM_BLOCKS
63 * Max number of devices supported by CUDA (for consistency checking).
65 * In reality it is 16 with CUDA <=v5.0, but let's stay on the safe side.
67 static int cuda_max_device_count = 32;
69 /** Dummy kernel used for sanity checking. */
70 __global__ void k_dummy_test()
75 /** Bit-flags which refer to memtestG80 test types and are used in do_memtest
76 * to specify which tests to run. */
77 enum memtest_G80_test_types {
78 MOVING_INVERSIONS_10 = 0x1,
79 MOVING_INVERSIONS_RAND = 0x2,
80 WALKING_8BIT_M86 = 0x4,
82 WALKING_1_8BIT = 0x10,
83 WALKING_0_32BIT = 0x20,
84 WALKING_1_32BIT = 0x40,
89 LOGIC_1_ITER_SHMEM = 0x800,
90 LOGIC_4_ITER_SHMEM = 0x1000
95 * \brief Runs GPU sanity checks.
97 * Runs a series of checks to determine that the given GPU and underlying CUDA
98 * driver/runtime functions properly.
99 * Returns properties of a device with given ID or the one that has
100 * already been initialized earlier in the case if of \dev_id == -1.
102 * \param[in] dev_id the device ID of the GPU or -1 if the device has already been initialized
103 * \param[out] dev_prop pointer to the structure in which the device properties will be returned
104 * \returns 0 if the device looks OK
106 * TODO: introduce errors codes and handle errors more smoothly.
108 static int do_sanity_checks(int dev_id, cudaDeviceProp *dev_prop)
113 cu_err = cudaGetDeviceCount(&dev_count);
114 if (cu_err != cudaSuccess)
116 fprintf(stderr, "Error %d while querying device count: %s\n", cu_err,
117 cudaGetErrorString(cu_err));
121 /* no CUDA compatible device at all */
127 /* things might go horribly wrong if cudart is not compatible with the driver */
128 if (dev_count < 0 || dev_count > cuda_max_device_count)
133 if (dev_id == -1) /* device already selected let's not destroy the context */
135 cu_err = cudaGetDevice(&id);
136 if (cu_err != cudaSuccess)
138 fprintf(stderr, "Error %d while querying device id: %s\n", cu_err,
139 cudaGetErrorString(cu_err));
146 if (id > dev_count - 1) /* pfff there's no such device */
148 fprintf(stderr, "The requested device with id %d does not seem to exist (device count=%d)\n",
154 memset(dev_prop, 0, sizeof(cudaDeviceProp));
155 cu_err = cudaGetDeviceProperties(dev_prop, id);
156 if (cu_err != cudaSuccess)
158 fprintf(stderr, "Error %d while querying device properties: %s\n", cu_err,
159 cudaGetErrorString(cu_err));
163 /* both major & minor is 9999 if no CUDA capable devices are present */
164 if (dev_prop->major == 9999 && dev_prop->minor == 9999)
168 /* we don't care about emulation mode */
169 if (dev_prop->major == 0)
176 cu_err = cudaSetDevice(id);
177 if (cu_err != cudaSuccess)
179 fprintf(stderr, "Error %d while switching to device #%d: %s\n",
180 cu_err, id, cudaGetErrorString(cu_err));
185 /* try to execute a dummy kernel */
186 k_dummy_test<<< 1, 512>>> ();
187 if (cudaThreadSynchronize() != cudaSuccess)
192 /* destroy context if we created one */
195 #if CUDA_VERSION < 4000
196 cu_err = cudaThreadExit();
197 CU_RET_ERR(cu_err, "cudaThreadExit failed");
199 cu_err = cudaDeviceReset();
200 CU_RET_ERR(cu_err, "cudaDeviceReset failed");
209 * \brief Runs a set of memory tests specified by the given bit-flags.
210 * Tries to allocate and do the test on \p megs Mb memory or
211 * the greatest amount that can be allocated (>10Mb).
212 * In case if an error is detected it stops without finishing the remaining
213 * steps/iterations and returns greater then zero value.
214 * In case of other errors (e.g. kernel launch errors, device querying errors)
217 * \param[in] which_tests variable with bit-flags of the requested tests
218 * \param[in] megs amount of memory that will be tested in MB
219 * \param[in] iter number of iterations
220 * \returns 0 if no error was detected, otherwise >0
222 static int do_memtest(unsigned int which_tests, int megs, int iter)
226 uint err_count; //, err_iter;
228 // no parameter check as this fn won't be called externally
230 // let's try to allocate the mem
231 while (!tester.allocate(megs) && (megs - 10 > 0))
233 megs -= 10; tester.deallocate();
238 fprintf(stderr, "Unable to allocate GPU memory!\n");
242 // clear the first 18 bits
243 which_tests &= 0x3FFF;
244 for (i = 0; i < iter; i++)
246 // Moving Inversions (ones and zeros)
247 if ((MOVING_INVERSIONS_10 & which_tests) == MOVING_INVERSIONS_10)
249 tester.gpuMovingInversionsOnesZeros(err_count);
252 return MOVING_INVERSIONS_10;
255 // Moving Inversions (random)
256 if ((MOVING_INVERSIONS_RAND & which_tests) == MOVING_INVERSIONS_RAND)
258 tester.gpuMovingInversionsRandom(err_count);
261 return MOVING_INVERSIONS_RAND;
264 // Memtest86 Walking 8-bit
265 if ((WALKING_8BIT_M86 & which_tests) == WALKING_8BIT_M86)
267 for (uint shift = 0; shift < 8; shift++)
269 tester.gpuWalking8BitM86(err_count, shift);
272 return WALKING_8BIT_M86;
276 // True Walking zeros (8-bit)
277 if ((WALKING_0_8BIT & which_tests) == WALKING_0_8BIT)
279 for (uint shift = 0; shift < 8; shift++)
281 tester.gpuWalking8Bit(err_count, false, shift);
284 return WALKING_0_8BIT;
288 // True Walking ones (8-bit)
289 if ((WALKING_1_8BIT & which_tests) == WALKING_1_8BIT)
291 for (uint shift = 0; shift < 8; shift++)
293 tester.gpuWalking8Bit(err_count, true, shift);
296 return WALKING_1_8BIT;
300 // Memtest86 Walking zeros (32-bit)
301 if ((WALKING_0_32BIT & which_tests) == WALKING_0_32BIT)
303 for (uint shift = 0; shift < 32; shift++)
305 tester.gpuWalking32Bit(err_count, false, shift);
308 return WALKING_0_32BIT;
312 // Memtest86 Walking ones (32-bit)
313 if ((WALKING_1_32BIT & which_tests) == WALKING_1_32BIT)
315 for (uint shift = 0; shift < 32; shift++)
317 tester.gpuWalking32Bit(err_count, true, shift);
320 return WALKING_1_32BIT;
325 if ((RANDOM_BLOCKS & which_tests) == RANDOM_BLOCKS)
327 tester.gpuRandomBlocks(err_count, rand());
330 return RANDOM_BLOCKS;
335 // Memtest86 Modulo-20
336 if ((MOD_20_32BIT & which_tests) == MOD_20_32BIT)
338 for (uint shift = 0; shift < 20; shift++)
340 tester.gpuModuloX(err_count, shift, rand(), 20, 2);
347 // Logic (one iteration)
348 if ((LOGIC_1_ITER & which_tests) == LOGIC_1_ITER)
350 tester.gpuShortLCG0(err_count, 1);
356 // Logic (4 iterations)
357 if ((LOGIC_4_ITER & which_tests) == LOGIC_4_ITER)
359 tester.gpuShortLCG0(err_count, 4);
366 // Logic (shared memory, one iteration)
367 if ((LOGIC_1_ITER_SHMEM & which_tests) == LOGIC_1_ITER_SHMEM)
369 tester.gpuShortLCG0Shmem(err_count, 1);
372 return LOGIC_1_ITER_SHMEM;
375 // Logic (shared-memory, 4 iterations)
376 if ((LOGIC_4_ITER_SHMEM & which_tests) == LOGIC_4_ITER_SHMEM)
378 tester.gpuShortLCG0Shmem(err_count, 4);
381 return LOGIC_4_ITER_SHMEM;
390 /*! \brief Runs a quick memory test and returns 0 in case if no error is detected.
391 * If an error is detected it stops before completing the test and returns a
392 * value greater then 0. In case of other errors (e.g. kernel launch errors,
393 * device querying errors) -1 is returned.
395 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
396 * \returns 0 if no error was detected, otherwise >0
398 int do_quick_memtest(int dev_id)
400 cudaDeviceProp dev_prop;
401 int devmem, res, time = 0;
405 time = getTimeMilliseconds();
408 if (do_sanity_checks(dev_id, &dev_prop) != 0)
410 // something went wrong
416 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
417 fprintf(debug, ">> Running QUICK memtests on %d MiB (out of total %d MiB), %d iterations\n",
418 QUICK_MEM, devmem, QUICK_ITER);
421 res = do_memtest(QUICK_TESTS, QUICK_MEM, QUICK_ITER);
425 fprintf(debug, "Q-RES = %d\n", res);
426 fprintf(debug, "Q-runtime: %d ms\n", getTimeMilliseconds() - time);
429 /* destroy context only if we created it */
437 /*! \brief Runs a full memory test and returns 0 in case if no error is detected.
438 * If an error is detected it stops before completing the test and returns a
439 * value greater then 0. In case of other errors (e.g. kernel launch errors,
440 * device querying errors) -1 is returned.
442 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
443 * \returns 0 if no error was detected, otherwise >0
446 int do_full_memtest(int dev_id)
448 cudaDeviceProp dev_prop;
449 int devmem, res, time = 0;
453 time = getTimeMilliseconds();
456 if (do_sanity_checks(dev_id, &dev_prop) != 0)
458 // something went wrong
462 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
466 fprintf(debug, ">> Running FULL memtests on %d MiB (out of total %d MiB), %d iterations\n",
467 devmem, devmem, FULL_ITER);
470 /* do all test on the entire memory */
471 res = do_memtest(FULL_TESTS, devmem, FULL_ITER);
475 fprintf(debug, "F-RES = %d\n", res);
476 fprintf(debug, "F-runtime: %d ms\n", getTimeMilliseconds() - time);
479 /* destroy context only if we created it */
487 /*! \brief Runs a time constrained memory test and returns 0 in case if no error is detected.
488 * If an error is detected it stops before completing the test and returns a value greater
489 * than zero. In case of other errors (e.g. kernel launch errors, device querying errors) -1
490 * is returned. Note, that test iterations are not interrupted therefor the total runtime of
491 * the test will always be multipple of one iteration's runtime.
493 * \param[in] dev_id the device id of the GPU or -1 if the device has laredy been selected
494 * \param[in] time_constr the time limit of the testing
495 * \returns 0 if no error was detected, otherwise >0
497 int do_timed_memtest(int dev_id, int time_constr)
499 cudaDeviceProp dev_prop;
500 int devmem, res = 0, time = 0, startt;
504 time = getTimeMilliseconds();
507 time_constr *= 1000; /* convert to ms for convenience */
508 startt = getTimeMilliseconds();
510 if (do_sanity_checks(dev_id, &dev_prop) != 0)
512 // something went wrong
516 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
520 fprintf(debug, ">> Running time constrained memtests on %d MiB (out of total %d MiB), time limit of %d s \n",
521 devmem, devmem, time_constr);
524 /* do the TIMED_TESTS set, one step at a time on the entire memory
525 that can be allocated, and stop when the given time is exceeded */
526 while ( ((int)getTimeMilliseconds() - startt) < time_constr)
528 res = do_memtest(TIMED_TESTS, devmem, 1);
537 fprintf(debug, "T-RES = %d\n", res);
538 fprintf(debug, "T-runtime: %d ms\n", getTimeMilliseconds() - time);
541 /* destroy context only if we created it */
549 /*! \brief Initializes the GPU with the given index.
551 * The varible \mygpu is the index of the GPU to initialize in the
552 * gpu_info.cuda_dev array.
554 * \param[in] mygpu index of the GPU to initialize
555 * \param[out] result_str the message related to the error that occurred
556 * during the initialization (if there was any).
557 * \param[in] gpu_info GPU info of all detected devices in the system.
558 * \param[in] gpu_opt options for using the GPUs in gpu_info
559 * \returns true if no error occurs during initialization.
561 gmx_bool init_gpu(int mygpu, char *result_str,
562 const gmx_gpu_info_t *gpu_info,
563 const gmx_gpu_opt_t *gpu_opt)
572 if (mygpu < 0 || mygpu >= gpu_opt->ncuda_dev_use)
574 sprintf(sbuf, "Trying to initialize an inexistent GPU: "
575 "there are %d %s-selected GPU(s), but #%d was requested.",
576 gpu_opt->ncuda_dev_use, gpu_opt->bUserSet ? "user" : "auto", mygpu);
580 gpuid = gpu_info->cuda_dev[gpu_opt->cuda_dev_use[mygpu]].id;
582 stat = cudaSetDevice(gpuid);
583 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
587 fprintf(stderr, "Initialized GPU ID #%d: %s\n", gpuid, gpu_info->cuda_dev[gpuid].prop.name);
590 return (stat == cudaSuccess);
593 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
595 * The context is explicitly destroyed and therefore all data uploaded to the GPU
596 * is lost. This should only be called when none of this data is required anymore.
598 * \param[out] result_str the message related to the error that occurred
599 * during the initialization (if there was any).
600 * \returns true if no error occurs during the freeing.
602 gmx_bool free_gpu(char *result_str)
611 stat = cudaGetDevice(&gpuid);
612 CU_RET_ERR(stat, "cudaGetDevice failed");
613 fprintf(stderr, "Cleaning up context on GPU ID #%d\n", gpuid);
616 #if CUDA_VERSION < 4000
617 stat = cudaThreadExit();
619 stat = cudaDeviceReset();
621 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
623 return (stat == cudaSuccess);
626 /*! \brief Returns true if the gpu characterized by the device properties is
627 * supported by the native gpu acceleration.
629 * \param[in] dev_prop the CUDA device properties of the gpus to test.
630 * \returns true if the GPU properties passed indicate a compatible
631 * GPU, otherwise false.
633 static bool is_gmx_supported_gpu(const cudaDeviceProp *dev_prop)
635 return (dev_prop->major >= 2);
638 /*! \brief Helper function that checks whether a given GPU status indicates compatible GPU.
640 * \param[in] stat GPU status.
641 * \returns true if the provided status is egpuCompatible, otherwise false.
643 static bool is_compatible_gpu(int stat)
645 return (stat == egpuCompatible);
648 /*! \brief Checks if a GPU with a given ID is supported by the native GROMACS acceleration.
650 * Returns a status value which indicates compatibility or one of the following
651 * errors: incompatibility, insistence, or insanity (=unexpected behavior).
652 * It also returns the respective device's properties in \dev_prop (if applicable).
654 * \param[in] dev_id the ID of the GPU to check.
655 * \param[out] dev_prop the CUDA device properties of the device checked.
656 * \returns the status of the requested device
658 static int is_gmx_supported_gpu_id(int dev_id, cudaDeviceProp *dev_prop)
663 stat = cudaGetDeviceCount(&ndev);
664 if (stat != cudaSuccess)
669 if (dev_id > ndev - 1)
671 return egpuNonexistent;
674 /* TODO: currently we do not make a distinction between the type of errors
675 * that can appear during sanity checks. This needs to be improved, e.g if
676 * the dummy test kernel fails to execute with a "device busy message" we
677 * should appropriately report that the device is busy instead of insane.
679 if (do_sanity_checks(dev_id, dev_prop) == 0)
681 if (is_gmx_supported_gpu(dev_prop))
683 return egpuCompatible;
687 return egpuIncompatible;
697 /*! \brief Detect all NVIDIA GPUs in the system.
699 * Will detect every NVIDIA GPU supported by the device driver in use. Also
700 * check for the compatibility of each and fill the gpu_info->cuda_dev array
701 * with the required information on each the device: ID, device properties,
704 * \param[in] gpu_info pointer to structure holding GPU information.
705 * \param[out] err_str The error message of any CUDA API error that caused
706 * the detection to fail (if there was any). The memory
707 * the pointer points to should be managed externally.
708 * \returns non-zero if the detection encountered a failure, zero otherwise.
710 int detect_cuda_gpus(gmx_gpu_info_t *gpu_info, char *err_str)
712 int i, ndev, checkres, retval;
715 cuda_dev_info_t *devs;
720 gpu_info->ncuda_dev_compatible = 0;
725 stat = cudaGetDeviceCount(&ndev);
726 if (stat != cudaSuccess)
730 /* cudaGetDeviceCount failed which means that there is something
731 * wrong with the machine: driver-runtime mismatch, all GPUs being
732 * busy in exclusive mode, or some other condition which should
733 * result in us issuing a warning a falling back to CPUs. */
735 s = cudaGetErrorString(stat);
736 strncpy(err_str, s, STRLEN*sizeof(err_str[0]));
741 for (i = 0; i < ndev; i++)
743 checkres = is_gmx_supported_gpu_id(i, &prop);
747 devs[i].stat = checkres;
749 if (checkres == egpuCompatible)
751 gpu_info->ncuda_dev_compatible++;
757 gpu_info->ncuda_dev = ndev;
758 gpu_info->cuda_dev = devs;
763 /*! \brief Select the GPUs compatible with the native GROMACS acceleration.
765 * This function selects the compatible gpus and initializes
766 * gpu_info->cuda_dev_use and gpu_info->ncuda_dev_use.
768 * Given the list of GPUs available in the system check each device in
769 * gpu_info->cuda_dev and place the indices of the compatible GPUs into
770 * cuda_dev_use with this marking the respective GPUs as "available for use."
771 * Note that \detect_cuda_gpus must have been called before.
773 * \param[in] gpu_info pointer to structure holding GPU information
774 * \param[in,out] gpu_opt pointer to structure holding GPU options
776 void pick_compatible_gpus(const gmx_gpu_info_t *gpu_info,
777 gmx_gpu_opt_t *gpu_opt)
783 /* cuda_dev/ncuda_dev have to be either NULL/0 or not (NULL/0) */
784 assert((gpu_info->ncuda_dev != 0 ? 0 : 1) ^ (gpu_info->cuda_dev == NULL ? 0 : 1));
786 snew(compat, gpu_info->ncuda_dev);
788 for (i = 0; i < gpu_info->ncuda_dev; i++)
790 if (is_compatible_gpu(gpu_info->cuda_dev[i].stat))
793 compat[ncompat - 1] = i;
797 gpu_opt->ncuda_dev_use = ncompat;
798 snew(gpu_opt->cuda_dev_use, ncompat);
799 memcpy(gpu_opt->cuda_dev_use, compat, ncompat*sizeof(*compat));
803 /*! \brief Check the existence/compatibility of a set of GPUs specified by their device IDs.
805 * Given the a list of gpu->ncuda_dev_use GPU device IDs stored in
806 * gpu_opt->cuda_dev_use check the existence and compatibility
807 * of the respective GPUs. Also provide the caller with an array containing
808 * the result of checks in \checkres.
810 * \param[out] checkres check result for each ID passed in \requested_devs
811 * \param[in] gpu_info pointer to structure holding GPU information
812 * \param[out] gpu_opt pointer to structure holding GPU options
813 * \returns TRUE if every the requested GPUs are compatible
815 gmx_bool check_selected_cuda_gpus(int *checkres,
816 const gmx_gpu_info_t *gpu_info,
817 gmx_gpu_opt_t *gpu_opt)
824 assert(gpu_opt->ncuda_dev_use >= 0);
826 if (gpu_opt->ncuda_dev_use == 0)
831 assert(gpu_opt->cuda_dev_use);
833 /* we will assume that all GPUs requested are valid IDs,
834 otherwise we'll bail anyways */
837 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
839 id = gpu_opt->cuda_dev_use[i];
841 /* devices are stored in increasing order of IDs in cuda_dev */
842 gpu_opt->cuda_dev_use[i] = id;
844 checkres[i] = (id >= gpu_info->ncuda_dev) ?
845 egpuNonexistent : gpu_info->cuda_dev[id].stat;
847 bAllOk = bAllOk && is_compatible_gpu(checkres[i]);
853 /*! \brief Frees the cuda_dev and cuda_dev_use array fields of \gpu_info.
855 * \param[in] gpu_info pointer to structure holding GPU information
857 void free_gpu_info(const gmx_gpu_info_t *gpu_info)
859 if (gpu_info == NULL)
864 sfree(gpu_info->cuda_dev);
867 /*! \brief Formats and returns a device information string for a given GPU.
869 * Given an index *directly* into the array of available GPUs (cuda_dev)
870 * returns a formatted info string for the respective GPU which includes
871 * ID, name, compute capability, and detection status.
873 * \param[out] s pointer to output string (has to be allocated externally)
874 * \param[in] gpu_info pointer to structure holding GPU information
875 * \param[in] index an index *directly* into the array of available GPUs
877 void get_gpu_device_info_string(char *s, const gmx_gpu_info_t *gpu_info, int index)
882 if (index < 0 && index >= gpu_info->ncuda_dev)
887 cuda_dev_info_t *dinfo = &gpu_info->cuda_dev[index];
890 dinfo->stat == egpuCompatible ||
891 dinfo->stat == egpuIncompatible;
895 sprintf(s, "#%d: %s, stat: %s",
897 gpu_detect_res_str[dinfo->stat]);
901 sprintf(s, "#%d: NVIDIA %s, compute cap.: %d.%d, ECC: %3s, stat: %s",
902 dinfo->id, dinfo->prop.name,
903 dinfo->prop.major, dinfo->prop.minor,
904 dinfo->prop.ECCEnabled ? "yes" : " no",
905 gpu_detect_res_str[dinfo->stat]);
909 /*! \brief Returns the device ID of the GPU with a given index into the array of used GPUs.
911 * Getter function which, given an index into the array of GPUs in use
912 * (cuda_dev_use) -- typically a tMPI/MPI rank --, returns the device ID of the
913 * respective CUDA GPU.
915 * \param[in] gpu_info pointer to structure holding GPU information
916 * \param[in] gpu_opt pointer to structure holding GPU options
917 * \param[in] idx index into the array of used GPUs
918 * \returns device ID of the requested GPU
920 int get_gpu_device_id(const gmx_gpu_info_t *gpu_info,
921 const gmx_gpu_opt_t *gpu_opt,
926 assert(idx >= 0 && idx < gpu_opt->ncuda_dev_use);
928 return gpu_info->cuda_dev[gpu_opt->cuda_dev_use[idx]].id;
931 /*! \brief Returns the device ID of the GPU currently in use.
933 * The GPU used is the one that is active at the time of the call in the active context.
935 * \param[in] gpu_info pointer to structure holding GPU information
936 * \returns device ID of the GPU in use at the time of the call
938 int get_current_gpu_device_id(void)
941 CU_RET_ERR(cudaGetDevice(&gpuid), "cudaGetDevice failed");
946 /*! \brief Returns the size of the cuda_dev_info struct.
948 * The size of cuda_dev_info can be used for allocation and communication.
950 * \returns size in bytes of cuda_dev_info
952 size_t sizeof_cuda_dev_info(void)
954 return sizeof(cuda_dev_info);